2D subspaces for sparse control of high-DOF robots

Area of Science:

• Computer Graphics
• Robotics
• Machine Learning

Background:

• Generating realistic hand movements is crucial for human-computer interaction and animation.
• Optical motion capture provides rich data but requires sophisticated processing for pose estimation.
• Existing methods often struggle with data sparsity and occlusions.

Purpose of the Study:

• To develop a low-dimensional (2D) parameterization of hand poses using manifold learning.
• To enable the generation of complex hand motions from sparse user input trajectories.
• To investigate the use of shape descriptors for handling occluded hand data.

Main Methods:

• Applied five dimension reduction and manifold learning techniques to optical motion capture data.
• Utilized shape descriptors to represent hand poses and address data occlusion.
• Developed a 2D parameterization for controlling hand movements.

Main Results:

• Successfully uncovered 2D parameterizations for power and precision grasps.
• Demonstrated the ability to drive a physically simulated hand using 2D mouse input.
• Showcased the potential for transforming sparse trajectories into desired hand movements.

Conclusions:

• Dimension reduction and manifold learning are effective for hand pose parameterization.
• The proposed 2D model facilitates intuitive control of simulated hands.
• This approach offers a promising direction for real-time hand motion generation.